Machine learning prediction of clinical tumor lysis syndrome in critically ill patients with hematologic malignancies using laboratory trajectory features: A MIMIC-IV retrospective cohort study.

263

Background: Tumor lysis syndrome (TLS) remains a life-threatening oncologic emergency in patients with hematologic malignancies admitted to the intensive care unit (ICU). Current risk stratification relies on static baseline factors such as malignancy type, white blood cell count, and lactate dehydrogenase (LDH) at presentation. Whether dynamic laboratory trajectories during the early ICU course can improve prediction of clinical TLS (CTLS) beyond baseline assessment alone has not been established. Methods: We conducted a retrospective cohort study using the MIMIC-IV database (v3.1). Adults (≥18 years) with high-risk hematologic malignancies, ICU admission, and complete TLS-defining laboratories (uric acid, potassium, phosphorus, calcium, creatinine) within 72 hours were included. Patients with pre-existing dialysis, prevalent TLS, or ICU stay <24 hours were excluded. CTLS was defined per Cairo-Bishop criteria. Trajectory features (slope, delta, peak, nadir, area under the curve) were extracted from serial measurements of seven analytes over 0–48 hours. A composite Metabolic Burden Score was derived from z-scores of peak uric acid, potassium, phosphorus, and nadir calcium. Three models (XGBoost, Random Forest, Logistic Regression) were trained on combined baseline and trajectory features and compared against a static-only baseline model. Model interpretability was assessed using SHAP analysis. Results: The final cohort comprised 415 patients (478 ICU stays) across eight malignancy subtypes. CTLS occurred in 129 patients (27.6%). CTLS was strongly associated with mortality (51.1% severe CTLS vs. 31.7% no TLS) and prolonged ICU stay (7.1 vs. 4.2 days). Patients with laboratory TLS alone had 0% mortality. Model performance and feature importance are shown in the Table. The trajectory-enhanced model improved AUC by 0.06 over static baseline alone (0.905 vs. 0.845). Three of the top five predictive features were dynamic trajectory variables, with Metabolic Burden Score ranking first (Table 1). Conclusions: In critically ill patients with hematologic malignancies, laboratory trajectory features derived from the first 48 hours of ICU admission substantially improve prediction of CTLS compared to static baseline risk factors. Dynamic monitoring of electrolyte and metabolite velocity may enable earlier identification of patients progressing toward clinical TLS.

Model performance and top predictive features.